1.5. Chaos—Or a Butterfly Spoils Laplace’s Dream 27 tiny force on Jupiter caused by a small asteroid had a significant ef- fect on its motion, then predicting the planetary orbits would be an impossible task. In the preceding section we saw that complete, C1 vector fields do give rise to a well-posed initial value problem, so Laplace seems to be on solid ground. Nevertheless, even though the initial value problems that arise in real-world applications may be technically well-posed in the above sense, they often behave as if they were ill-posed. For a class of examples that turns up frequently—the so-called chaotic systems— predictability is only an unachievable theoretical ideal. While their short-term behavior is predictable, on longer time-scales prediction becomes, for practical purposes, impossible. This may seem para- doxical at first if we have an algorithm for predicting accurately for ten seconds, then should not repeating it with that first prediction as a new initial condition provide an accurate prediction for twenty seconds? Unfortunately, a hallmark feature of chaotic systems, called “sensitive dependence on initial conditions”, defeats this strategy. Let us consider an initial value problem dx dt = V (x), x(0) = p0 and see how things go wrong for a chaotic system when we try to compute σp 0 (t) for large t. Suppose that p1 is very close to p0, say p0 − p1 δ, and let us compare σp 1 (t) and σp 0 (t). Continuity with respect to initial conditions tells us that for δ small enough σp 1 (t) at least initially will not diverge too far from σp 1 (t). In fact, for a chaotic system, a typical behavior—when p0 is near a so-called “strange attractor”—is for σp 1 (t) to at first “track” σp 0 (t) in the sense that σp 0 (t) − σp 1 (t) initially stays nearly constant or even decreases—so in particular the motions of σp 0 (t) and σp 1 (t) are highly correlated. But then, suddenly, there will be a period during which σp 1 (t) starts to veer off in a different direction, following which σp 0 (t) − σp 1 (t) will grow exponentially fast for a while. Soon they will be far apart, and although their distance remains bounded, from that time forward their motions become completely uncorrelated. If we make δ smaller, then we can guarantee that σp 1 (t) will track σp 0 (t) for a longer period, but (and this is the essence of sensitive dependence on initial condi- tions) no matter how small we make δ, the veering away and loss of correlation will always occur. The reason this is relevant

Purchased from American Mathematical Society for the exclusive use of nofirst nolast (email unknown) Copyright 2009 American Mathematical Society. Duplication prohibited. Please report unauthorized use to cust-serv@ams.org. Thank You! Your purchase supports the AMS' mission, programs, and services for the mathematical community.